A Comparison of Convolutional Neural Networks and Feature-based Machine Learning Methods for the Ripeness Classification of Strawberries

Abstract

A variety of machine learning methods are often used for ripeness detection of fruits and vegetables using image data. Existing research in this area often focuses only on training feature-based classifiers or on using raw images with convolutional neural networks. The purpose of this paper is to compare both approaches in terms of their classification accuracy. To answer our research question, we analyze the performance of convolutional neural networks and different feature-based classifiers on a balanced dataset consisting of three strawberry ripeness classes: unripe, ripe, and overripe. Our investigation shows that convolutional neural networks outperform almost all feature-based classifier. However, the penalized multinomial regression achieves the best accuracy of 86.27 % without any hyper-parameter tuning. Another insight is that different methods lead to the best sensitivity for different ripeness classes. Convolutional neural networks most accurately classify unripe strawberries, while ripe strawberries are best classified by penalized discriminant analysis and overripe berries are best classified by penalized multinomial regression.